The performance of single-cell foundation models is dependent on many factors. This study assesses the effect of the pretraining dataset’s size and diversity, revealing potential challenges in pursuing consistent improvement by naively scaling up pretraining data.